Amol Kapoor says HTML beats canvas for AI agent graphics

We spend countless hours dragging bounding boxes, snapping to grids, and aligning pixels in design software. This geospatial workflow suits human eyes and hands perfectly. But when we build tools like Figma Model Context Protocols or PowerPoint command-line interfaces for AI agents, we force them into our own physical constraints. The result is almost always a chaotic, misaligned mess.

The fatal flaw of SVG and coordinate systems

AI skeptics point to benchmarks like Arc AGI to argue that neural networks cannot reason about space. This limitation becomes glaringly obvious in the famous "pelican on a bicycle" test popularized by developer Simon Willison. Ask a model to draw this scene using Scalable Vector Graphics, and you will get a horrifying cluster of disconnected shapes. The breakdown does not happen because the model lacks creativity; it happens because SVG is essentially a wall of coordinate numbers. Humans cannot write raw SVG by hand, and models struggle to map abstract token predictions to precise pixel placement without feedback loops.

Why HTML is the native graphic language for LLMs

The fix is simple: stop treating the model like a human designer. Instead of coordinate grids, give the AI a medium built on structure and semantics. HTML is the ideal candidate. Models have trained on billions of lines of web markup. They understand the relationship between a header, a section, and a flexbox container. By letting the browser handle the actual rendering, the agent focuses on structural logic rather than absolute positioning.

Generating production decks and video from structural code

Using this approach, Amol Kapoor and his team at Nori build board decks, sales materials, and entire videos using nothing but structured markup. This shifts the developer's role from fixing visual bugs to managing data flow. By feeding the model internal business data, team members can assemble slide decks directly from Slack history or code repositories. Stop thinking like a software user and start designing tools that match how large language models actually process information.